我正在尝试使用Lambda层在Keras中构建裁剪模块，但是在尝试编译模型时，我始终收到以下错误消息：
“ AttributeError：NoneType对象没有属性'_inbound_nodes'”
这是我要完成的工作：
模型1的输出为None，6、34、60张量，它是对象位置热图的粗略近似值。
coarse_model = load_model('coarse_model.h5')
input_1 = Input(shape=(270,480,3))
input_2 = Input(shape=(135,240,3))
output_hmap = coarse_model([input_1,input_2])
在这里，我从原始模型中提取了一些中间层，并将其用作裁剪的输入：
intermediate_layer_model1 = coarse_model.get_layer("input1").output
intermediate_layer_model2 = coarse_model.get_layer("input2").output
intermediate_layer_model3 = coarse_model.get_layer("input3").output
intermediate_layer_model4 = coarse_model.get_layer("input4").output
第1层的形状为None，270,480,128，并且随着您进入更深的层而逐渐变小。这是我要裁剪的图层，围绕先前模型的输出位置放置了一个36x36的补丁。裁剪模块很复杂，是Keras中的Lambda层。我试图将其纯粹用张量流编写，并且期望的输出是每个None批次的裁剪部分，每个intermediate_layer中的6层都是如此。
layer1 = Lambda(crop_module,output_shape=(36,36,128))
([output_hmap,K.constant(36),tf.constant(0,dtype=tf.int32),
K.constant([270,480],dtype=tf.float32),intermediate_layer_model1])
def crop_module(data):
output = data[0]
dimension = data[1]
idx = data[2]
shaped = data[3]
def hmap_to_coord(hmap):
# flatten the Tensor along the height and width axes
flat_tensor = tf.cast(flat()(hmap),dtype='float32')
index_y = tf.argmax(flat_tensor,axis=0)
# argmax of the flat tensor
index_y = tf.cast(tf.divide(tf.cast(tf.argmax(flat_tensor,axis=1),tf.float32),(tf.cast(tf.shape(hmap)[2],tf.float32))),tf.int32)
index_x = tf.argmax(flat_tensor,axis=1,output_type=tf.int32) - index_y*tf.shape(hmap)[2]
index_y = tf.cast(tf.divide(index_y+1,tf.shape(hmap)[1]),tf.float32)
index_x = tf.cast(tf.divide(index_x+1,tf.shape(hmap)[2]),tf.float32)
# stack and return 2D coordinates
return tf.stack([index_x,index_y])
def find_coordinates(coordinates,dim,shape):
shape = tf.cast(shape,tf.float32)
dim = tf.constant(36)
half_dim = tf.cast(tf.divide(dim,2),dtype=tf.int32)
i = tf.constant(0)
steps = tf.shape(coordinates)[0]
initial_outputs = tf.TensorArray(dtype=tf.float32, size=steps)
def while_condition(i,*args):
return tf.less(i, steps)
def body(i,shape,outputs_):
# do something here which you want to do in your loop
# increment i
coordinates_x = tf.cast(tf.multiply(shape[1],coordinates[i][0]),tf.int32)
coordinates_y = tf.cast(tf.multiply(coordinates[i][1],shape[0]),tf.int32)
shape = tf.cast(shape,tf.int32)
x = tf.cond(tf.less(coordinates_x,dim+1),lambda: tf.stack([0,shape[1] - dim ]),
lambda: tf.cond(tf.less(coordinates_x, shape[1] - dim+1), lambda: tf.stack((shape[1] - dim,0)),
lambda: tf.cond(tf.equal(dim,9),lambda:tf.stack((coordinates_x - 5,shape[1] - coordinates_x - 4)),
lambda:tf.stack((coordinates_x - half_dim,shape[1] - coordinates_x - half_dim)))))
y = tf.cond(tf.less(coordinates_y,dim+1),lambda: tf.stack([shape[0] - dim, 0 ]),
lambda: tf.cond(tf.less(coordinates_y, shape[1] - dim+1), lambda: tf.stack((0,shape[0] - dim)),
lambda: tf.cond(tf.equal(dim,9),lambda:tf.stack((shape[0] - coordinates_y - 4,coordinates_y-5)),
lambda:tf.stack((shape[0] - coordinates_y - half_dim,coordinates_y-half_dim)))))
outs = tf.stack([x,y])
outputs_.write(i,outs)
return tf.add(i, 1),shape,outputs_
# do the loop:
r,shape,output = tf.while_loop(while_condition, body, [i,shape,initial_outputs])
return output.stack()
### Get the coordinates for each point from the hmap output of model1
output = tf.cast(output,tf.int32)
idx = tf.cast(idx,tf.int32)
coordinat = hmap_to_coord(output[:,idx])
coord = tf.cast(find_coordinates(coordinat,dimension,shaped),tf.int32)
## do this for each of the points in the output
i = tf.constant(0)
steps = tf.shape(coordinat)[0]
initial_outputs = tf.TensorArray(dtype=tf.float32, size=steps)
def while_condition(i,*args):
return tf.less(i, steps)
def body(i,data,coord1,shaped,outputs_):
# do something here which you want to do in your loop
# increment i
shaped = tf.cast(shaped,tf.int32)
crop1 = data[4][i][coord1[i][0][0]:shaped[0]-coord1[i][0][1],coord1[i][1][0]:shaped[1]-coord1[i][1][1],:]
outputs_.write(i,crop1)
return tf.add(i, 1),data,coord1,shaped,outputs_
# do the loop:
r,dat,cord1,shape,outputs = tf.while_loop(while_condition, body, [i,data,coord,shaped,initial_outputs])
return outputs.stack()
如果我尝试编译模型，即使是在第一层，也会出现inbound_nodes错误。
final_model = Model(inputs=[input_1,input_2],outputs=layer1)
我是否可能在Keras中实现目标？能否通过其他模型的张量确定每批的裁剪尺寸？